@debugg-ai/debugg-ai-mcp 3.3.0 → 3.5.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -5,6 +5,45 @@ All notable changes to the DebuggAI MCP project will be documented in this file.
5
5
  The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
6
6
  and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
7
7
 
8
+ ## [3.5.0]
9
+
10
+ ### Added — Remote transport: Streamable HTTP + OAuth Resource Server (opt-in)
11
+
12
+ The server can now run as a hosted, multi-user remote MCP over **stateless
13
+ Streamable HTTP**, in addition to the default stdio transport (which is
14
+ unchanged). Enable with `DEBUGGAI_MCP_TRANSPORT=http` (+ `PORT`, default 3000).
15
+
16
+ As an OAuth **Resource Server** (MCP 2025-06-18):
17
+ - Each `POST /mcp` request must carry `Authorization: Bearer <token>`; the token
18
+ is request-scoped (AsyncLocalStorage) and used as the backend credential —
19
+ `api.debugg.ai` is the validator, so no token-verification keys live here.
20
+ - Missing/invalid token → `401` with `WWW-Authenticate: Bearer resource_metadata=…`.
21
+ - Serves RFC 9728 metadata at `/.well-known/oauth-protected-resource` advertising
22
+ the authorization server (`auth.debugg.ai`), so clients run the OAuth flow and
23
+ retry with a token.
24
+ - `GET /health` for load-balancer / ECS health checks.
25
+
26
+ Auth became request-scoped without touching the ~20 backend-client call sites:
27
+ `config.api.key` resolves the per-request token when set (`utils/requestContext.ts`).
28
+ Config env: `DEBUGGAI_MCP_TRANSPORT`, `PORT`, `DEBUGGAI_MCP_PUBLIC_URL`,
29
+ `DEBUGGAI_OAUTH_ISSUER`, `DEBUGGAI_TOKEN_TYPE=bearer`. stdio installs need none of these.
30
+
31
+ ## [3.4.0]
32
+
33
+ ### Added — MCP Resources (browse projects / environments / executions)
34
+
35
+ The server now declares the `resources` capability and exposes the read-only
36
+ entities as addressable resources, so clients can browse and @-mention them as
37
+ context instead of only calling tools:
38
+
39
+ - **Collections** (`resources/list`): `debugg-ai://projects`, `debugg-ai://environments`, `debugg-ai://executions`
40
+ - **Templates** (`resources/templates/list`): `debugg-ai://project/{uuid}`, `debugg-ai://environment/{uuid}`, `debugg-ai://execution/{uuid}`
41
+ - **`resources/read`** dispatches each URI to the same entity handler the tools
42
+ use — identical data + auth, no drift — and returns the JSON payload.
43
+
44
+ Additive: clients without resource support keep using the tools unchanged.
45
+ Implementation in `handlers/resourcesHandler.ts`.
46
+
8
47
  ## [3.3.0]
9
48
 
10
49
  ### Added — Run artifacts returned as resource links
package/README.md CHANGED
@@ -158,6 +158,24 @@ Every filter-mode response is paginated. Response shape:
158
158
 
159
159
  Pass optional `page` (1-indexed, default 1) and `pageSize` (default 20, max 200; oversized values are clamped). No response is ever silently truncated.
160
160
 
161
+ ## Resources
162
+
163
+ Alongside tools, the server exposes the read-only entities as MCP **resources**
164
+ so clients can browse and @-mention them as context:
165
+
166
+ | URI | What |
167
+ |---|---|
168
+ | `debugg-ai://projects` | All projects (first page) |
169
+ | `debugg-ai://environments` | Environments for the auto-detected project |
170
+ | `debugg-ai://executions` | Recent executions (first page) |
171
+ | `debugg-ai://project/{uuid}` | One project, full detail |
172
+ | `debugg-ai://environment/{uuid}` | One environment (credentials inline, passwords redacted) |
173
+ | `debugg-ai://execution/{uuid}` | One execution, full node detail + artifact links |
174
+
175
+ Reads dispatch to the same handlers as the `project` / `environment` /
176
+ `executions` tools, so the data and auth are identical. Resources are additive —
177
+ clients without resource support keep using the tools.
178
+
161
179
  ### Security invariants
162
180
 
163
181
  - Passwords are write-only. They never appear in any response body from any tool.
@@ -216,6 +234,37 @@ Response-shape changes: the bare `count` field on list responses is gone — use
216
234
  DEBUGGAI_API_KEY=your_api_key
217
235
  ```
218
236
 
237
+ ## Remote / HTTP transport (optional)
238
+
239
+ By default the server speaks **stdio** (local `npx`). It can instead run as a
240
+ hosted, multi-user remote MCP over **stateless Streamable HTTP** + OAuth:
241
+
242
+ ```bash
243
+ DEBUGGAI_MCP_TRANSPORT=http PORT=3000 DEBUGGAI_TOKEN_TYPE=bearer npx -y @debugg-ai/debugg-ai-mcp@latest
244
+ ```
245
+
246
+ It is an OAuth **Resource Server**: every `POST /mcp` needs
247
+ `Authorization: Bearer <token>`; missing/invalid tokens get a `401` with a
248
+ `WWW-Authenticate` pointing at the RFC 9728 metadata, and clients run the OAuth
249
+ flow against the advertised authorization server. The bearer is request-scoped —
250
+ `api.debugg.ai` validates it.
251
+
252
+ | Endpoint | Purpose |
253
+ |---|---|
254
+ | `POST /mcp` | MCP Streamable HTTP (bearer-protected) |
255
+ | `GET /.well-known/oauth-protected-resource` | RFC 9728 metadata (authorization server discovery) |
256
+ | `GET /health` | Load-balancer / ECS health check |
257
+
258
+ | Env var | Default | Purpose |
259
+ |---|---|---|
260
+ | `DEBUGGAI_MCP_TRANSPORT` | `stdio` | Set to `http` for the remote transport |
261
+ | `PORT` | `3000` | HTTP listen port |
262
+ | `DEBUGGAI_MCP_PUBLIC_URL` | `https://mcp.debugg.ai` | This server's public resource URL (RFC 9728 `resource`) |
263
+ | `DEBUGGAI_OAUTH_ISSUER` | `https://auth.debugg.ai` | Authorization server advertised to clients |
264
+ | `DEBUGGAI_TOKEN_TYPE` | `token` | Set to `bearer` so OAuth tokens forward as `Authorization: Bearer` |
265
+
266
+ stdio installs need none of these.
267
+
219
268
  ## Telemetry
220
269
 
221
270
  The MCP server ships with telemetry enabled by default — an embedded write-only PostHog project key (`phc_*`) so the team can observe cache hit rates, poll cadence, tunnel reliability, and other operational metrics across the install base. Captured events:
@@ -5,6 +5,7 @@ import { z } from 'zod';
5
5
  import { readFileSync } from 'fs';
6
6
  import { fileURLToPath } from 'url';
7
7
  import { dirname, join } from 'path';
8
+ import { currentApiKey } from '../utils/requestContext.js';
8
9
  function findPackageVersion() {
9
10
  const __dir = dirname(fileURLToPath(import.meta.url));
10
11
  let dir = __dir;
@@ -113,7 +114,14 @@ let _config;
113
114
  export const config = {
114
115
  get server() { return getConfig().server; },
115
116
  get devMode() { return getConfig().devMode; },
116
- get api() { return getConfig().api; },
117
+ // api.key is request-scoped under the HTTP transport: if a per-request token
118
+ // is set (AsyncLocalStorage), it overrides the env key for that request only.
119
+ // stdio / tests have no request store, so the env key is returned unchanged.
120
+ get api() {
121
+ const api = getConfig().api;
122
+ const requestKey = currentApiKey();
123
+ return requestKey ? { ...api, key: requestKey } : api;
124
+ },
117
125
  get defaults() { return getConfig().defaults; },
118
126
  get logging() { return getConfig().logging; },
119
127
  get telemetry() { return getConfig().telemetry; },
@@ -0,0 +1,127 @@
1
+ /**
2
+ * MCP Resources (epic pglam).
3
+ *
4
+ * Exposes the read-only entities (projects / environments / executions) as
5
+ * addressable resources so clients can browse and @-mention them as context
6
+ * instead of (or alongside) calling the `project`/`environment`/`executions`
7
+ * tools. Reads reuse the exact tool handlers — same data, same auth, no drift.
8
+ *
9
+ * Collections (resources/list): debugg-ai://projects | environments | executions
10
+ * Items (resources/templates/list): debugg-ai://project/{uuid}
11
+ * debugg-ai://environment/{uuid}
12
+ * debugg-ai://execution/{uuid}
13
+ *
14
+ * Resources are additive: clients that don't support the capability simply
15
+ * keep using the tools.
16
+ */
17
+ import { config } from '../config/index.js';
18
+ import { MCPError, MCPErrorCode, } from '../types/index.js';
19
+ import { projectHandler } from './projectHandler.js';
20
+ import { environmentHandler } from './environmentHandler.js';
21
+ import { executionsHandler } from './executionsHandler.js';
22
+ const SCHEME = 'debugg-ai';
23
+ const JSON_MIME = 'application/json';
24
+ /** Concrete collection resources returned by resources/list. */
25
+ export const RESOURCE_COLLECTIONS = [
26
+ {
27
+ uri: `${SCHEME}://projects`,
28
+ name: 'projects',
29
+ title: 'DebuggAI Projects',
30
+ description: 'All projects visible to this API key (first page).',
31
+ mimeType: JSON_MIME,
32
+ },
33
+ {
34
+ uri: `${SCHEME}://environments`,
35
+ name: 'environments',
36
+ title: 'DebuggAI Environments',
37
+ description: 'Environments for the auto-detected project (credentials redacted).',
38
+ mimeType: JSON_MIME,
39
+ },
40
+ {
41
+ uri: `${SCHEME}://executions`,
42
+ name: 'executions',
43
+ title: 'DebuggAI Executions',
44
+ description: 'Recent workflow executions (first page).',
45
+ mimeType: JSON_MIME,
46
+ },
47
+ ];
48
+ /** URI templates returned by resources/templates/list. */
49
+ export const RESOURCE_TEMPLATES = [
50
+ {
51
+ uriTemplate: `${SCHEME}://project/{uuid}`,
52
+ name: 'project',
53
+ title: 'DebuggAI Project',
54
+ description: 'A single project by UUID, with full detail.',
55
+ mimeType: JSON_MIME,
56
+ },
57
+ {
58
+ uriTemplate: `${SCHEME}://environment/{uuid}`,
59
+ name: 'environment',
60
+ title: 'DebuggAI Environment',
61
+ description: 'A single environment by UUID, with credentials inline (passwords redacted).',
62
+ mimeType: JSON_MIME,
63
+ },
64
+ {
65
+ uriTemplate: `${SCHEME}://execution/{uuid}`,
66
+ name: 'execution',
67
+ title: 'DebuggAI Execution',
68
+ description: 'A single execution by UUID, with full node detail + artifact links.',
69
+ mimeType: JSON_MIME,
70
+ },
71
+ ];
72
+ function readContext() {
73
+ return { timestamp: new Date() };
74
+ }
75
+ /** Pull the JSON payload that every entity handler emits as its single text block. */
76
+ function payloadText(res) {
77
+ const text = res.content?.find((c) => c.type === 'text')?.text;
78
+ return typeof text === 'string' ? text : '{}';
79
+ }
80
+ // debugg-ai://<kind> (collection)
81
+ // debugg-ai://<kind>/<uuid> (item)
82
+ const URI_RE = new RegExp(`^${SCHEME}://([a-z]+)(?:/([^/?#]+))?$`);
83
+ /**
84
+ * Resolve a debugg-ai:// resource URI by dispatching to the matching tool
85
+ * handler and wrapping its JSON payload as a resource content block.
86
+ */
87
+ export async function readResource(uri) {
88
+ if (!config.api.key) {
89
+ throw new MCPError(MCPErrorCode.CONFIGURATION_ERROR, 'DEBUGGAI_API_KEY is not set. Configure it in your MCP server registration. Get a key at https://debugg.ai.', { missingEnvVars: ['DEBUGGAI_API_KEY'] });
90
+ }
91
+ const match = URI_RE.exec(uri);
92
+ if (!match) {
93
+ throw new MCPError(MCPErrorCode.INVALID_PARAMS, `Unrecognized resource URI: ${uri}`);
94
+ }
95
+ const [, kind, id] = match;
96
+ const ctx = readContext();
97
+ const requireId = () => {
98
+ if (!id) {
99
+ throw new MCPError(MCPErrorCode.INVALID_PARAMS, `Resource ${uri} requires a UUID: ${SCHEME}://${kind}/{uuid}`);
100
+ }
101
+ return id;
102
+ };
103
+ let res;
104
+ switch (kind) {
105
+ case 'projects':
106
+ res = await projectHandler({ action: 'list' }, ctx);
107
+ break;
108
+ case 'project':
109
+ res = await projectHandler({ action: 'get', uuid: requireId() }, ctx);
110
+ break;
111
+ case 'environments':
112
+ res = await environmentHandler({ action: 'list' }, ctx);
113
+ break;
114
+ case 'environment':
115
+ res = await environmentHandler({ action: 'get', uuid: requireId() }, ctx);
116
+ break;
117
+ case 'executions':
118
+ res = await executionsHandler({ action: 'list' }, ctx);
119
+ break;
120
+ case 'execution':
121
+ res = await executionsHandler({ action: 'get', uuid: requireId() }, ctx);
122
+ break;
123
+ default:
124
+ throw new MCPError(MCPErrorCode.INVALID_PARAMS, `Unknown resource kind "${kind}" in ${uri}`);
125
+ }
126
+ return { contents: [{ uri, mimeType: JSON_MIME, text: payloadText(res) }] };
127
+ }
@@ -0,0 +1,134 @@
1
+ /**
2
+ * Streamable HTTP transport + OAuth Resource Server (epic lybfq).
3
+ *
4
+ * Opt-in remote transport: `DEBUGGAI_MCP_TRANSPORT=http` (stdio stays default).
5
+ * Stateless (no session id) so it scales behind a plain load balancer.
6
+ *
7
+ * Auth model — the MCP server is an OAuth **Resource Server**:
8
+ * - Every /mcp request must carry `Authorization: Bearer <token>`.
9
+ * - The token is stashed per-request (AsyncLocalStorage) and used as the
10
+ * backend credential; api.debugg.ai is the real validator (a bad token 401s
11
+ * on the first backend call). No token verification keys live here.
12
+ * - Missing token → 401 + `WWW-Authenticate: Bearer resource_metadata=...`,
13
+ * and we serve RFC 9728 metadata at /.well-known/oauth-protected-resource
14
+ * pointing clients at auth.debugg.ai to run the OAuth flow.
15
+ *
16
+ * Deployment note: set DEBUGGAI_TOKEN_TYPE=bearer so the backend client forwards
17
+ * the OAuth token as `Authorization: Bearer` (not `Token`).
18
+ */
19
+ import { createServer } from 'node:http';
20
+ import { StreamableHTTPServerTransport } from '@modelcontextprotocol/sdk/server/streamableHttp.js';
21
+ import { runWithApiKey } from './utils/requestContext.js';
22
+ const PUBLIC_URL = (process.env.DEBUGGAI_MCP_PUBLIC_URL || 'https://mcp.debugg.ai').replace(/\/+$/, '');
23
+ const OAUTH_ISSUER = (process.env.DEBUGGAI_OAUTH_ISSUER || 'https://auth.debugg.ai').replace(/\/+$/, '');
24
+ const RESOURCE_METADATA_PATH = '/.well-known/oauth-protected-resource';
25
+ const MCP_PATH = '/mcp';
26
+ const MAX_BODY_BYTES = 8 * 1024 * 1024;
27
+ /** RFC 9728 protected-resource metadata: tells clients which AS issues tokens. */
28
+ export function protectedResourceMetadata() {
29
+ return {
30
+ resource: PUBLIC_URL,
31
+ authorization_servers: [OAUTH_ISSUER],
32
+ bearer_methods_supported: ['header'],
33
+ };
34
+ }
35
+ /** Extract the token from `Authorization: Bearer <t>` (or `Token <t>`). */
36
+ export function bearerToken(authHeader) {
37
+ if (!authHeader)
38
+ return undefined;
39
+ const m = /^(?:Bearer|Token)\s+(.+)$/i.exec(authHeader.trim());
40
+ return m ? m[1].trim() : undefined;
41
+ }
42
+ function sendJson(res, code, body, extraHeaders = {}) {
43
+ const data = JSON.stringify(body);
44
+ res.writeHead(code, {
45
+ 'Content-Type': 'application/json',
46
+ 'Content-Length': Buffer.byteLength(data),
47
+ ...extraHeaders,
48
+ });
49
+ res.end(data);
50
+ }
51
+ function unauthorized(res) {
52
+ const metadataUrl = `${PUBLIC_URL}${RESOURCE_METADATA_PATH}`;
53
+ sendJson(res, 401, { error: 'unauthorized', error_description: 'Missing or invalid bearer token; authenticate via the linked authorization server.' }, { 'WWW-Authenticate': `Bearer resource_metadata="${metadataUrl}"` });
54
+ }
55
+ function readJsonBody(req) {
56
+ return new Promise((resolve, reject) => {
57
+ let data = '';
58
+ let aborted = false;
59
+ req.on('data', (chunk) => {
60
+ data += chunk;
61
+ if (data.length > MAX_BODY_BYTES && !aborted) {
62
+ aborted = true;
63
+ reject(new Error('request body too large'));
64
+ }
65
+ });
66
+ req.on('end', () => {
67
+ if (aborted)
68
+ return;
69
+ if (!data)
70
+ return resolve(undefined);
71
+ try {
72
+ resolve(JSON.parse(data));
73
+ }
74
+ catch (e) {
75
+ reject(e);
76
+ }
77
+ });
78
+ req.on('error', reject);
79
+ });
80
+ }
81
+ /** Start the stateless Streamable HTTP server. Resolves to the listening server. */
82
+ export async function startHttpServer(opts) {
83
+ const { port, buildServer, logger } = opts;
84
+ const httpServer = createServer(async (req, res) => {
85
+ const path = new URL(req.url || '/', 'http://localhost').pathname;
86
+ // ECS / LB health check — no auth.
87
+ if (path === '/health' && req.method === 'GET') {
88
+ return sendJson(res, 200, { status: 'ok' });
89
+ }
90
+ // RFC 9728 protected-resource metadata — public discovery, no auth.
91
+ if (path === RESOURCE_METADATA_PATH && req.method === 'GET') {
92
+ return sendJson(res, 200, protectedResourceMetadata());
93
+ }
94
+ if (path === MCP_PATH) {
95
+ const token = bearerToken(req.headers['authorization']);
96
+ if (!token) {
97
+ logger.info('HTTP MCP request without bearer token → 401');
98
+ return unauthorized(res);
99
+ }
100
+ let body;
101
+ if (req.method === 'POST') {
102
+ try {
103
+ body = await readJsonBody(req);
104
+ }
105
+ catch {
106
+ return sendJson(res, 400, { error: 'invalid_request', error_description: 'Request body must be valid JSON' });
107
+ }
108
+ }
109
+ // Stateless: a fresh server + transport per request, scoped to this
110
+ // request's bearer token via AsyncLocalStorage (so config.api.key resolves
111
+ // to it for every backend call made while handling this request).
112
+ const srv = buildServer();
113
+ const transport = new StreamableHTTPServerTransport({ sessionIdGenerator: undefined });
114
+ res.on('close', () => {
115
+ transport.close().catch(() => { });
116
+ srv.close().catch(() => { });
117
+ });
118
+ try {
119
+ await srv.connect(transport);
120
+ await runWithApiKey(token, () => transport.handleRequest(req, res, body));
121
+ }
122
+ catch (error) {
123
+ logger.error('HTTP MCP request failed', { error: error instanceof Error ? error.message : String(error) });
124
+ if (!res.headersSent)
125
+ sendJson(res, 500, { error: 'internal_error' });
126
+ }
127
+ return;
128
+ }
129
+ sendJson(res, 404, { error: 'not_found' });
130
+ });
131
+ await new Promise((resolve) => httpServer.listen(port, resolve));
132
+ logger.info('HTTP transport listening', { port, resource: PUBLIC_URL, authorizationServer: OAUTH_ISSUER });
133
+ return httpServer;
134
+ }
package/dist/index.js CHANGED
@@ -18,9 +18,10 @@
18
18
  */
19
19
  import { Server } from "@modelcontextprotocol/sdk/server/index.js";
20
20
  import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
21
- import { CallToolRequestSchema, ListToolsRequestSchema, } from "@modelcontextprotocol/sdk/types.js";
21
+ import { CallToolRequestSchema, ListToolsRequestSchema, ListResourcesRequestSchema, ListResourceTemplatesRequestSchema, ReadResourceRequestSchema, } from "@modelcontextprotocol/sdk/types.js";
22
22
  import { config } from "./config/index.js";
23
23
  import { initTools, getTools, getTool } from "./tools/index.js";
24
+ import { readResource, RESOURCE_COLLECTIONS, RESOURCE_TEMPLATES } from "./handlers/resourcesHandler.js";
24
25
  import { Logger, validateInput, createErrorResponse, toMCPError, handleConfigurationError, Telemetry, TelemetryEvents, withStructuredContent, } from "./utils/index.js";
25
26
  import { MCPErrorCode, MCPError, } from "./types/index.js";
26
27
  // Logger and server are initialized lazily in main() to avoid triggering
@@ -39,18 +40,21 @@ function createMCPServer() {
39
40
  tools: {
40
41
  listChanged: false,
41
42
  },
43
+ resources: {
44
+ listChanged: false,
45
+ },
42
46
  },
43
47
  });
44
48
  }
45
49
  /**
46
50
  * Create progress callback for tool execution
47
51
  */
48
- function createProgressCallback(progressToken) {
52
+ function createProgressCallback(srv, progressToken) {
49
53
  if (!progressToken)
50
54
  return undefined;
51
55
  return async ({ progress, total, message }) => {
52
56
  try {
53
- await server.notification({
57
+ await srv.notification({
54
58
  method: "notifications/progress",
55
59
  params: {
56
60
  progressToken,
@@ -72,10 +76,21 @@ function createProgressCallback(progressToken) {
72
76
  };
73
77
  }
74
78
  /**
75
- * Register MCP request handlers. Called in main() after server is created.
79
+ * Build a fully-configured MCP Server (capabilities + all request handlers).
80
+ * The HTTP transport calls this once per request for stateless isolation; main()
81
+ * uses it for the long-lived stdio server.
82
+ */
83
+ export function buildConfiguredServer() {
84
+ const srv = createMCPServer();
85
+ registerHandlers(srv);
86
+ return srv;
87
+ }
88
+ /**
89
+ * Register MCP request handlers on a server instance. Called for the stdio
90
+ * singleton in main(), and once per request by the HTTP transport (stateless).
76
91
  */
77
- function registerHandlers() {
78
- server.setRequestHandler(CallToolRequestSchema, async (req) => {
92
+ function registerHandlers(srv) {
93
+ srv.setRequestHandler(CallToolRequestSchema, async (req) => {
79
94
  const typedReq = req;
80
95
  const requestId = `req_${Date.now()}`;
81
96
  const requestLogger = logger.child({ requestId });
@@ -109,7 +124,7 @@ function registerHandlers() {
109
124
  requestId,
110
125
  timestamp: new Date(),
111
126
  };
112
- const progressCallback = createProgressCallback(typeof progressToken === 'string' || typeof progressToken === 'number' ? String(progressToken) : undefined);
127
+ const progressCallback = createProgressCallback(srv, typeof progressToken === 'string' || typeof progressToken === 'number' ? String(progressToken) : undefined);
113
128
  requestLogger.info(`Executing tool: ${name}`);
114
129
  const toolStart = Date.now();
115
130
  const result = await tool.handler(validatedInput, context, progressCallback);
@@ -130,11 +145,33 @@ function registerHandlers() {
130
145
  return createErrorResponse(mcpError, typedReq.params.name);
131
146
  }
132
147
  });
133
- server.setRequestHandler(ListToolsRequestSchema, async () => {
148
+ srv.setRequestHandler(ListToolsRequestSchema, async () => {
134
149
  const tools = getTools();
135
150
  logger.info('Tools list requested', { toolCount: tools.length });
136
151
  return { tools };
137
152
  });
153
+ // Resources (epic pglam): browse projects/environments/executions as
154
+ // addressable read URIs. Reads dispatch to the same entity handlers as the
155
+ // tools, so data + auth stay consistent.
156
+ srv.setRequestHandler(ListResourcesRequestSchema, async () => {
157
+ logger.info('Resources list requested', { resourceCount: RESOURCE_COLLECTIONS.length });
158
+ return { resources: RESOURCE_COLLECTIONS };
159
+ });
160
+ srv.setRequestHandler(ListResourceTemplatesRequestSchema, async () => {
161
+ return { resourceTemplates: RESOURCE_TEMPLATES };
162
+ });
163
+ srv.setRequestHandler(ReadResourceRequestSchema, async (req) => {
164
+ const uri = req.params?.uri;
165
+ logger.info('Resource read requested', { uri });
166
+ try {
167
+ return await readResource(uri);
168
+ }
169
+ catch (error) {
170
+ const mcpError = toMCPError(error, 'resource read');
171
+ logger.error('Resource read failed', { uri, errorCode: mcpError.code, message: mcpError.message });
172
+ throw mcpError;
173
+ }
174
+ });
138
175
  }
139
176
  /**
140
177
  * Main server initialization and startup
@@ -144,19 +181,18 @@ async function main() {
144
181
  // Initialize logger and server here (not at module load time) so config
145
182
  // validation errors are caught by this try-catch instead of crashing.
146
183
  logger = new Logger({ module: 'main' });
147
- server = createMCPServer();
148
- // Register request handlers (they reference the `server` variable)
149
- registerHandlers();
184
+ const transportMode = (process.env.DEBUGGAI_MCP_TRANSPORT || 'stdio').toLowerCase();
150
185
  logger.info('Starting DebuggAI MCP Server', {
151
186
  nodeVersion: process.version,
152
187
  platform: process.platform,
153
188
  architecture: process.arch,
154
- pid: process.pid
189
+ pid: process.pid,
190
+ transport: transportMode,
155
191
  });
156
- // NOTE: DEBUGGAI_API_KEY validation is deferred to the first tool call so
157
- // MCP clients see a proper initialize response + a structured tool error,
158
- // rather than the subprocess dying with "Failed to reconnect" (bead cma).
159
- if (!config.api.key) {
192
+ // stdio is single-user: the API key comes from the environment and is
193
+ // validated at first tool call (bead cma). HTTP is multi-user: each request
194
+ // carries its own bearer token, so a missing env key at boot is expected.
195
+ if (transportMode !== 'http' && !config.api.key) {
160
196
  logger.warn('DEBUGGAI_API_KEY is not set. Server will boot but every tool call will return a ConfigurationError until the env var is configured.');
161
197
  }
162
198
  // Initialize telemetry: PostHog by default (public project key embedded
@@ -180,12 +216,28 @@ async function main() {
180
216
  // No API calls at boot. Project context is resolved lazily on first tool
181
217
  // invocation (list_environments / list_credentials / check_app_in_browser).
182
218
  initTools(null);
183
- const transport = new StdioServerTransport();
184
- await server.connect(transport);
185
- logger.info('DebuggAI MCP Server is running and ready to accept requests', {
186
- transport: 'stdio',
187
- toolsAvailable: getTools().map(t => t.name),
188
- });
219
+ if (transportMode === 'http') {
220
+ // Remote/hosted transport (epic lybfq): stateless Streamable HTTP + OAuth
221
+ // Resource Server. stdio stays the default and is unaffected.
222
+ const { startHttpServer } = await import('./httpServer.js');
223
+ const port = Number(process.env.PORT) || 3000;
224
+ await startHttpServer({ port, buildServer: buildConfiguredServer, logger });
225
+ logger.info('DebuggAI MCP Server is running and ready to accept requests', {
226
+ transport: 'http',
227
+ port,
228
+ toolsAvailable: getTools().map(t => t.name),
229
+ });
230
+ }
231
+ else {
232
+ server = createMCPServer();
233
+ registerHandlers(server);
234
+ const transport = new StdioServerTransport();
235
+ await server.connect(transport);
236
+ logger.info('DebuggAI MCP Server is running and ready to accept requests', {
237
+ transport: 'stdio',
238
+ toolsAvailable: getTools().map(t => t.name),
239
+ });
240
+ }
189
241
  }
190
242
  catch (error) {
191
243
  logger.error('Failed to start DebuggAI MCP Server', {
@@ -0,0 +1,24 @@
1
+ /**
2
+ * Request-scoped context (epic lybfq).
3
+ *
4
+ * stdio is single-user: one API key from the environment for the whole process.
5
+ * The HTTP transport is multi-user: each request carries its own bearer token.
6
+ * Rather than thread a token through every handler + backend-client call site,
7
+ * we stash it in AsyncLocalStorage for the duration of the request and let
8
+ * `config.api.key` resolve it (see config/index.ts). Outside an HTTP request
9
+ * (i.e. stdio, or tests) the store is empty and the env key is used — so the
10
+ * stdio path is completely unchanged.
11
+ *
12
+ * This module intentionally has NO imports so any layer (incl. config) can use
13
+ * it without creating an import cycle.
14
+ */
15
+ import { AsyncLocalStorage } from 'node:async_hooks';
16
+ const storage = new AsyncLocalStorage();
17
+ /** Run `fn` with a request-scoped API key (used by the HTTP transport per request). */
18
+ export function runWithApiKey(apiKey, fn) {
19
+ return storage.run({ apiKey }, fn);
20
+ }
21
+ /** The request-scoped API key if inside runWithApiKey(), else undefined (stdio). */
22
+ export function currentApiKey() {
23
+ return storage.getStore()?.apiKey;
24
+ }
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@debugg-ai/debugg-ai-mcp",
3
- "version": "3.3.0",
3
+ "version": "3.5.0",
4
4
  "description": "Zero-Config, Fully AI-Managed End-to-End Testing for all code gen platforms.",
5
5
  "type": "module",
6
6
  "bin": {